It is genuinely strange to watch the same company get expelled from one customer and onboarded by every other customer in the same six-week window. That is the position Anthropic is in right now, and the strangeness is the story, because it is not random.
The split
On one side: the Pentagon designated Anthropic a “supply chain risk” in early March, which is a bureaucratic instrument normally reserved for Chinese telecom equipment makers and shell companies in Cyprus. Defense Secretary Pete Hegseth was visibly the driver. The $200 million contract Anthropic signed last July is now winding down on a six-month phase-out ordered by President Trump via executive order. A California federal judge ruled the designation violated Anthropic’s First Amendment and due-process rights. The DC Circuit declined to stay the designation pending appeal. Oral arguments happened May 19. A ruling is pending. In the meantime, the Pentagon stood up GenAI.mil and put 25 “power users” on OpenAI, Google, and Grok to replace Claude inside classified workflows.
On the other side: Microsoft finalized the Foundry model catalog at 11,000+ entries with Claude Opus 4.8, Sonnet 4.5, and Haiku 4.5 stocked as first-class options alongside GPT-5.5 and Gemini. Apple is reportedly making Claude one of three selectable AI Extensions in iOS 27, alongside Gemini (the default) and ChatGPT, which would make Claude a native option on roughly two billion devices the moment iOS 27 ships in September. Anthropic’s web-traffic share grew from 203 million visits in January to 824 million in April, a 306 percent quarterly jump that puts it at 8.2 percent worldwide and 12.5 percent in the US. Cognition raised at a $26 billion valuation on a coding agent built on Claude. Anthropic itself closed its own round at $9.65 billion in May.
If you graphed Anthropic’s standing across customer segments right now, the chart would look broken.
Why one customer is different
The split is not noise. The Pentagon is the only customer in Anthropic’s universe that explicitly asked Anthropic to drop its two stated red lines: no fully autonomous weapons that target humans without a human in the loop, and no mass surveillance of Americans. Anthropic kept both. Hegseth got on television. The supply-chain-risk designation followed. The court fight now playing out in DC is, stripped of procedural surface area, about whether a frontier-model lab is allowed to refuse a government use case without being banned for the refusal.
No other Anthropic customer has asked them to cross either line, which is why no other Anthropic customer has a complaint. Microsoft is selling Claude to enterprise buyers who want a model with good coding performance and visible safety posture. Apple wants a model their users can opt into without Apple having to own its outputs. Coding-agent startups want a model that handles long context and reasoning well. None of those use cases test the red lines, so the red lines are invisible there. The Pentagon use case tested them on day one.
What the market is voting on
Look at the integrations as votes. Each one is a major buyer saying “we want this model in our stack, with its safety policies attached, including the ones we will never personally hit.” The buyers are not naive. Microsoft and Apple know exactly what Anthropic’s policies say, because the policies are public and the policies are the brand. Those buyers chose to take the policy attached.
That is a useful tell about where the enterprise market is on AI safety in 2026. The conventional wisdom inside frontier labs has been that “safety theater” is a drag on commercial uptake and the buyers who actually pay want fewer guardrails, faster. The Anthropic integration sweep is the counter-evidence. The buyers who pay want guardrails they can point to in board decks. They want the lab’s name attached. They want a deployment story where, if a model does something stupid in production, “we chose the safety-positioned lab” is a credible defense for the procurement lead. Anthropic is, in that sense, selling reputational insurance bundled with reasoning capacity.
What the Pentagon fight actually decides
If the DC Circuit upholds the supply-chain-risk designation, the precedent is genuinely chilling for the rest of the industry, because it means any lab can be expelled from federal procurement for refusing a use case. Every other frontier lab is therefore watching this case with the kind of attention they would prefer not to show, because if Anthropic loses, the next time a lab declines a national-security request, the same playbook runs against them. If the DC Circuit reverses and sides with the California court, then the precedent is roughly “labs are allowed to enforce written safety policies and the federal government has to either accept those policies or buy from someone else.” Either outcome reshapes federal AI procurement.
The current set of facts suggests Anthropic priced this correctly. They got beaten up in Washington, then absorbed by everyone outside Washington at a faster rate than they were losing share in Washington. That math, more than anything else, is why the company is raising at a $9.65 billion valuation while a sitting Defense Secretary is trying to ban them. It turns out the market for “the lab that holds its lines” was larger than the market for “the lab that doesn’t.”